Loading ToolBench
MCP server that lets any MCP-compatible AI agent delegate to Claude Code (Opus 4.7 by default) via your Claude Max subscription OAuth — not API billing.
Single tool 'ask_claude' has a well-written description (234 chars, within 10-1024 baseline) that clearly states WHAT it does, WHEN to use it, and prerequisites. Input schema is present with proper type and description. However, the tool lacks output schema documentation, error handling guidance, and parameter constraints. The description does not mention that this is a stateful delegation (no memory of calling agent's conversation) until mid-sentence, which could confuse LLM selection. No pagination, no structured output schema, no recovery guidance for failures.
Delegate a substantive task to Claude Code, billed against the user's Claude Max / Pro subscription (model: claude-opus-4-7). Use this for: code generation, deep reasoning, multi-step engineering, refactoring, or anything that needs more capability than the routing model can reliably deliver. The receiver has NO memory of the calling agent's conversation — include everything Claude needs in the prompt argument: the user's verbatim request, project / file context, relevant prior messages, and what kind of output you want back. Returns Claude's full text response.
No output schema documented. Tool returns text response but LLM cannot infer structure, field names, or how to chain downstream calls.
No error handling guidance. If claude CLI fails, times out, or returns non-zero exit, the error message is raw stderr. LLM cannot determine if error is retryable, user-fixable, or fatal.
Prompt parameter lacks format/length constraints. Description says 'include all context' but does not specify max length, encoding, or what happens if prompt exceeds token limits.
Tool name 'ask_claude' is generic and does not follow verb_noun pattern (e.g., 'delegate_to_claude' or 'invoke_claude_code' would be clearer). 'ask' is vague about what happens.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
No idempotency guarantee documented. If an agent retries after a timeout, will the same prompt be re-executed, potentially causing duplicate work or side effects?